CAIS proposes a safety agenda for an AI slowdown

CAIS outlines work it argues could fill an AI slowdown: stronger containment, less harmful AI behavior, defenses against jailbreaks and prompt injection, and better government capacity to manage AI. It distinguishes capabilities—what AI can do—from propensities—what it tends to do—and calls for reducing tendencies to lie, cheat or cause harm. The organization also advocates more independent AI safety funders, exploration of alternative safety approaches, and targeted post-training data for beneficial uses such as radiology, weather forecasting and agriculture.